A Data-Driven Newsvendor Problem : Insights from Machine Learning and Feature

Analytically solving the classic newsvendor problem relies on a known demand distribution. In practice, however, it is not easy to approximate a sufficiently accurate demand distribution. This results in the obtained analytical solution may not be optimal. To address this issue, we propose a novel data-driven newsvendor model using machine learning algorithms that can incorporate demand features into replenishment decisions for solving single- and multi-product newsvendor problems; in particular, our algorithms do not require the demand distribution and only rely on historical data. With the extracted key features of demand information from historical data, our algorithms can simultaneously optimize demand estimation and replenishment decisions. First, we propose a frequency alignment method to transform high-dimensional mixed-frequency data and historical data into the same frequency; this can effectively avoid losing the demand i nformation. Second, two variable selection policies are proposed based on the empirical risk minimization principle to tackle the parameter proliferation issue; we numerically justify that the optimal order quantity can be obtained with the minimum total cost. Third, a feature-based machine learning algorithm is also designed for solving the multi-product newsvendor problem. We find that using real-time information to update high-dimensional mixed-frequency data can effectively reduce the total c ost. Numerical experiments validate that the identified key features of demand can be sufficiently exploited to make data-informed replenishment decisions. Results show that the optimal replenishment quantity behaves robustly and does not fluctuate greatly given different cost st ructures. Our work provides meaningful insights for newsvendors making replenishment decisions under stochastic demand